Open Source OpenClaw Alternatives

Hermes Agent is a desktop AI agent built by Nous Research that runs natively on macOS, Windows, and Linux. It's designed for people who want a single agent that works across their communication tools, handles recurring tasks, and actually remembers what it's done before.

The core idea is persistence. Most agents start fresh every session. Hermes maintains memory across conversations, auto-generates skills from past interactions, and builds up context about your projects over time. You don't have to re-explain things.

Where it connects:

• Telegram, Discord, Slack, WhatsApp, Signal, and email are all supported natively, so the agent meets you where you already work
• A CLI interface is also available for terminal-first workflows

What it can do:

• Natural-language scheduling for reports, backups, and briefings that run unattended
• Web search and browser automation with vision, image generation, and text-to-speech
• Subagents that run in isolated conversations with their own terminals and Python RPC scripts, useful for parallel pipelines without ballooning context costs
• Multi-model reasoning with access to 300+ models through the Nous Portal

Sandboxing is a genuine differentiator. Five execution backends (local, Docker, SSH, Singularity, and Modal) give you control over where code actually runs, with container hardening and namespace isolation. That matters if you're running automated tasks or untrusted scripts.

For teams or individuals already using LobeChat or similar multi-surface chat interfaces, Hermes takes a different angle: it's less about chat UI and more about an agent that operates autonomously across surfaces. Think of it as closer to an AgentOS approach, where the agent itself is the persistent layer.

It's free to use under the MIT license, with paid tiers through the Nous Portal for higher model access and monthly credits.

NanoClaw is a self-hosted personal AI agent built for individuals who want full control over their AI assistant. It connects to messaging apps including WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, WeChat, and email, then routes every conversation through isolated Docker containers. The codebase is intentionally small: 132 source files, roughly 17,500 lines of code, and fewer than 10 dependencies. You can read the whole thing in an afternoon.

It positions itself as a lightweight alternative to OpenClaw, which ships 3,680 source files and 70 dependencies. That size difference isn't cosmetic. It shapes whether you can audit what your agent actually does, customize it without fear, and trust its security model.

Security is structural, not policy-based. Each agent group runs in its own Linux container with its own filesystem. It can only see directories you explicitly mount. Credentials never enter the container at all. Outbound API requests route through OneCLI's Agent Vault, which injects authentication at the proxy level and enforces per-agent rate limits and policies.

Key capabilities:

• Multi-channel messaging – WhatsApp, Telegram, Slack, Discord, Teams, iMessage, Matrix, and more, installed on demand with /add-<channel> skills
• Flexible agent wiring – give each channel its own isolated agent, share one agent across channels for unified memory, or fold channels into a shared session
• Per-agent workspaces – each agent group has its own memory, its own CLAUDE.md, and its own container boundary
• Scheduled tasks – recurring jobs that run Claude and message you back (morning briefings, weekly reviews)
• Multiple AI providers – runs Claude Code natively via the Claude Agent SDK; drop-in options for OpenAI Codex, OpenRouter, Google, DeepSeek, and local models via Ollama
• Skills over features – install only the adapters you need; nothing is bundled that you didn't ask for

The architecture is a single Node.js host process that routes inbound messages through an entity model, writes to SQLite, and wakes per-session containers. No microservices, no message brokers, no shared memory across agent boundaries.

NanoClaw is MIT-licensed and designed to be forked. The philosophy is that your personal AI agent should be working software shaped to your exact needs, not a generic framework you configure around.

Eigent is a desktop application that puts a multi-agent AI workforce on your machine. You describe a task, and a pool of specialized agents divides it up, works in parallel, and delivers results. It's built for people who want AI to actually finish work, not just assist with it.

The core idea is "cowork": agents operate alongside you in your desktop environment with access to your local files, browser, and terminal. You can run a single focused agent for direct tasks or spin up a full workforce where agents collaborate across steps. A Browser Agent handles web research, a Document Agent reads and writes files, a Terminal Agent runs shell commands, and a Multi-modal Agent processes images and other media.

Key capabilities include:

• Scheduled automation: set recurring workflows so agents run tasks on a timer, even when you're not at your desk
• Local-first execution: files, credentials, and context stay on your machine; nothing has to leave it
• Model agnostic: connect cloud APIs, enterprise model gateways, or local inference. No vendor lock-in
• MCP and skill support: track which tools, skills, and referenced files each agent uses during a task
• Agent folder: every file an agent creates or modifies during a task is surfaced in one place for review

The open source version is fully self-hostable with your own API keys or local models. It's built on CAMEL-AI, a multi-agent framework, and runs on Mac, Windows, and Linux.

For teams doing research-heavy work, competitive analysis, report generation, or any task that involves pulling from multiple sources and producing structured output, Eigent is closer to OpenHands than a simple chat interface. It's designed to complete multi-step work end-to-end rather than hand control back to you at every turn.

OpenBot is a self-hosted enterprise agent platform built for companies that can't afford to hand their data to a third party. It runs on your own infrastructure, keeps documents and conversations in your own Postgres database, and enforces the same file permissions your existing systems already have. Every agent gets its own channel, behaves like a named colleague, and only has access to the tools and sources its role requires.

The platform ships several agent archetypes out of the box:

• Knowledge agent answers questions from connected company documents (Google Drive, OneDrive) and always cites the source files, so answers can be verified rather than trusted.
• Metrics agent queries your analytics warehouse and draws the answer directly instead of describing it in prose.
• Research agent gathers context across documents and drafts from them, keeping contradictory sources rather than quietly picking one.
• Tool operator signs into interfaces that have no useful API and drives them the way a person would. You can watch the screen in real time and take control mid-task.
• Inbox triage, onboarding buddy, and other role-based agents can be published and scoped to exactly the skills they need.

Permissions are fail-closed throughout. An agent only sees files the person asking is already allowed to open. Documents with ambiguous permission mappings aren't returned at all. Sensitive tool calls go through a gateway with an explicit per-tool allowlist, and every call is logged.

OpenBot is built on AG-UI, the Agent-User Interaction Protocol that the team behind OpenBot wrote and maintains. This means any agent that speaks AG-UI, whether built on LangGraph, Mastra, PydanticAI, or a custom framework, can be registered as an endpoint and given a channel. No lock-in to a single agent framework.

On the infrastructure side: no shared tenancy, no seat counting, and no bundled model. Admins supply their own provider credentials, which are encrypted at rest and never logged. Nothing silently falls back to an external LLM if a provider isn't configured. Helm is the supported production deployment path.

Rakazo is a self-hosted platform for running AI agents that actually do things. Each bot gets a sandboxed browser and shell, signs into your tools the way you would, and completes tasks while you're away. It's a direct Grok Bot alternative built around one principle: you own everything.

Bots run in Docker on your own machine, behind your firewall. Sessions and credentials never leave your infrastructure. You point each bot at whatever model you want (Claude, GPT, Grok, or a local model), and you can mix and match per bot. The cheap model triages; the smart one writes.

Setup starts with an interview. A new bot asks you a few questions about the work, your writing style, and which tools it should use. Then it gets going.

What bots can handle out of the box:

• Inbox Manager archives noise, replies to routine threads, and parks drafts for your review
• Sales Outbound researches accounts overnight, scores intent, and drafts outreach in your voice
• Talent Scout reads applicants, shortlists against your criteria, and writes intro emails
• Bug Triage reproduces reports in a real browser and attaches reproduction steps to the issue
• Expense Manager matches receipts to charges and files reports, asking before guessing
• Chief of Staff runs briefings, bookings, and handoffs between your other bots

Routines are saved as plain Markdown. Show a bot a workflow once and it writes a routine you can read, edit, and commit to version control. Approvals are configurable: set what a bot may do alone and what it must ask you about first. Every action lands in an audit log you own.

Rakazo is Apache-2.0 licensed with no seats, no pricing tiers, and nothing gated. Unlimited bots, no limits on usage. A managed cloud option (where you bring your own keys and they run the sandboxes) is planned, with no migration required when it arrives.